AI Reads Routine Ultrasound Scans to Predict Breast Cancer Spread to Lymph Nodes

For women diagnosed with invasive breast cancer, one of the most consequential questions after the initial tumor is found is deceptively simple: has the cancer reached the lymph nodes under the arm? The answer shapes surgery, staging, and the entire treatment plan, yet today it usually requires a surgical procedure called sentinel lymph node biopsy, in which surgeons remove the first draining node and pathologists examine it under a microscope. A new study published in BMC Medical Imaging suggests that a deep learning system trained on ordinary preoperative ultrasound images may one day help answer that question before a patient ever enters the operating room, offering a noninvasive way to rank patients by their risk of nodal spread.

The research, led by Kun He, Yansheng Qiu, and Baohui Zeng with senior author Xinmin Guo, was conducted as a retrospective two-center study involving 246 patients with pathologically confirmed invasive breast cancer. The team drew its training data from Guangzhou Red Cross Hospital of Jinan University, splitting 194 patients from that institution into a training cohort of 155 and an internal test cohort of 39. A further 52 patients from Guangdong Provincial Hospital of Integrated Traditional Chinese and Western Medicine served as an external validation cohort, a critical step because models that perform well on the data they were trained on often falter when confronted with images from unfamiliar scanners, protocols, and patient populations.

The technical heart of the study lies in what the authors call deep learning-derived radiomics. Traditional radiomics relies on handcrafted mathematical features, such as texture, shape, and intensity statistics, extracted from medical images by predefined formulas. Deep learning takes a different route: a convolutional neural network, in this case the well-known ResNet50 architecture, ingests raw ultrasound images and learns its own hierarchical set of features, from simple edges and echogenicity patterns in early layers to increasingly abstract representations of tumor appearance in deeper layers. These learned features can capture subtle image characteristics that human-designed metrics might miss, and that even experienced sonographers may not consciously perceive.

In practice, the researchers fed preoperative ultrasound images of the primary breast tumors into ResNet50, harvested the resulting deep features, and then applied a feature selection procedure within the training cohort to whittle down the high-dimensional feature space. After selection, 22 deep learning-derived features were retained. These were used to train a gradient boosting decision tree classifier, an ensemble machine learning method that builds many sequential decision trees, each one correcting the errors of its predecessors, to produce a probability that a given patient’s sentinel lymph node harbors metastasis. The pipeline also incorporated standard tools of the radiomics trade, including least absolute shrinkage and selection operator regularization for feature reduction, intraclass correlation coefficients to assess feature reproducibility, and Shapley additive explanations to interpret which features drove the model’s predictions.

Performance was assessed with receiver operating characteristic analysis, the standard method for evaluating how well a model separates positive from negative cases across all decision thresholds. The model achieved an area under the curve of 0.734 in the training cohort, 0.761 in the internal test cohort, and 0.742 in the external validation cohort. An AUC of 0.5 represents random guessing, while 1.0 represents perfect discrimination, so values in the low-to-mid 0.7 range indicate moderate, but far from definitive, discriminatory power. Notably, the point estimates remained similar across all three cohorts, which is encouraging for generalizability, since a model that collapses when moved to a new hospital would be of little practical value.

In the external validation cohort, the model reached an accuracy of 0.769, with sensitivity of 0.650, specificity of 0.844, a positive predictive value of 0.722, and a negative predictive value of 0.794. Translated into clinical terms, the model was better at correctly identifying patients whose nodes were cancer-free than at catching every patient with metastasis. Its sensitivity of 65 percent means roughly one in three patients with nodal spread would be missed, a limitation the authors acknowledge directly. The relatively stronger specificity suggests the model could be most useful for identifying low-risk patients who might safely avoid more extensive axillary procedures, rather than for ruling out metastasis with certainty.

The authors are unusually candid about the limits of their work. The confidence intervals around the AUC estimates are wide, spanning for example 0.598 to 0.886 in the external cohort, reflecting the modest sample sizes involved, particularly the 52-patient external set. Decision-curve analysis was performed only as an exploratory evaluation of net benefit, and the study explicitly states that the decision curves do not establish clinical utility. The conclusion emphasizes that the model provides preliminary discrimination and risk ranking only, is not currently clinically actionable, and requires prospective validation in larger cohorts before any clinical use. This restraint matters in a field where artificial intelligence claims often outpace the evidence supporting them.

Why does preoperative prediction of sentinel lymph node status matter so much? The sentinel node is the first lymph node to receive drainage from the tumor, and its status is one of the strongest prognostic factors in early breast cancer. Current practice relies on sentinel lymph node biopsy, which, while far less morbid than the full axillary dissections of the past, still carries risks of lymphedema, numbness, seroma, and infection. If a reliable imaging-based predictor could identify patients with a very low probability of nodal involvement, it could potentially spare some patients the biopsy altogether, or conversely flag high-risk patients for more thorough axillary evaluation and neoadjuvant treatment planning. Ultrasound is an especially attractive platform because it is inexpensive, widely available, radiation-free, and already part of routine breast cancer workup in most of the world, meaning the data needed to run such a model already exists in the clinical record.

The study also illustrates a broader trend in medical imaging research: the marriage of deep neural networks with radiomics-style pipelines and interpretable machine learning classifiers. Rather than asking a neural network to make the final diagnosis end to end, the authors used ResNet50 as a feature extractor and then applied a gradient boosting decision tree on top, an approach that can work well with limited training data and allows the use of established interpretability tools such as SHAP values to reveal which learned image features influence predictions. As imaging datasets grow and multi-center collaborations mature, this hybrid strategy may yield models that are both more accurate and more transparent than purely black-box alternatives.

For now, the message for patients and clinicians is one of cautious optimism. The two-center results show that information about axillary metastasis is genuinely encoded in the pixel patterns of routine breast ultrasound, and that a machine learning pipeline can extract enough of it to achieve moderate discrimination across different hospitals. But with wide confidence intervals, modest sensitivity, and no prospective testing, the model remains a research tool rather than a clinical one. The authors’ call for larger, prospective validation cohorts is the necessary next step, and if those studies confirm and strengthen the current findings, ultrasound-based deep learning radiomics could eventually become a standard, noninvasive component of individualized axillary management in invasive breast cancer, reducing surgical burden for low-risk patients while helping direct intensive treatment to those who need it most.

Subject of Research: Deep learning-derived ultrasound radiomics for preoperative prediction of sentinel lymph node metastasis in invasive breast cancer

Article Title: Ultrasound-based deep learning-derived radiomics for preoperative prediction of sentinel lymph node metastasis in invasive breast cancer: a two-center retrospective study

Article References: He, K., Qiu, Y., Zeng, B., Huang, Z., Lin, J., Chen, J., Mei, Y., Lai, J., & Guo, X. (2026). Ultrasound-based deep learning-derived radiomics for preoperative prediction of sentinel lymph node metastasis in invasive breast cancer: a two-center retrospective study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02919-7

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02919-7

Keywords: breast cancer, sentinel lymph node, ultrasound, deep learning, radiomics, ResNet50, machine learning, metastasis prediction, external validation, cancer imaging, axillary management, predictive medicine